AStatisticalApproachtoNeuralNetworksforPatternRecognition

Anaccessibleandup-to-datetreatmentfeaturingtheconnectionbetweenneuralnetworksandstatisticsAStatisticalApproachtoNeuralNetworksforPatternRecognitionpresentsastatisticaltreatmentoftheMultilayerPerceptron(MLP),whichisthemostwidelyusedoftheneuralnetworkmodels.Thisbookaimstoanswerquestionsthatarisewhenstatisticiansarefirstconfrontedwiththistypeofmodel,suchas:Howrobustisthemodeltooutliers?Couldthemodelbemademorerobust?Whichpointswillhaveahighleverage?Whataregoodstartingvaluesforthefittingalgorithm?Thoroughanswerstothesequestionsandmanymoreareincluded,aswellasworkedexamplesandselectedproblemsforthereader.DiscussionsontheuseofMLPmodelswithspatialandspectraldataarealsoincluded.FurthertreatmentofhighlyimportantprincipalaspectsoftheMLPareprovided,suchastherobustnessofthemodelintheeventofoutlyingoratypicaldata;theinfluenceandsensitivitycurvesoftheMLP;whytheMLPisafairlyrobustmodel;andmodificationstomaketheMLPmorerobust.Theauthoralsoprovidesclarificationofseveralmisconceptionsthatareprevalentinexistingneuralnetworkliterature.Throughoutthebook,theMLPmodelisextendedinseveraldirectionstoshowthatastatisticalmodelingapproachcanmakevaluablecontributions,andfurtherexplorationforfittingMLPmodelsismadepossibleviatheRandS-PLUS®codesthatareavailableonthebook'srelatedWebsite.AStatisticalApproachtoNeuralNetworksforPatternRecognitionsuccessfullyconnectslogisticregressionandlineardiscriminantanalysis,thusmakingitacriticalreferenceandself-studyguideforstudentsandprofessionalsalikeinthefieldsofmathematics,statistics,computerscience,andelectricalengineering.